A roadmap is only useful if it is honest about tense. Most infrastructure roadmaps blur the line between what is running, what is being built, and what someone hopes to build someday, and the result is a document that a prospective customer cannot act on.
This one keeps the three apart. Everything below is sorted into what the Cobble Network does today, what is in active development with engineering effort behind it, and what is a longer-term ambition that depends on partnerships, capital, or work not yet started. Each item is tied to the part of the mission it serves: making inference more sustainable, more resilient, or more accessible. Where an item is already covered in depth elsewhere on this site, it is summarized rather than repeated.
Completed: what runs today
An open-weight model catalog served at production quality. The network serves a curated set of open-weight models for chat, reasoning, coding, vision, and embeddings through an OpenAI-compatible endpoint. Every listed model passed conformance testing for tool calling, structured output, prompt caching, and streaming before it was published. Accessible.
A fleet organized by architecture, on reclaimed hardware. Capacity is grouped into pools by model architecture, each running the serving engine and weight format that suits it, on equipment reclaimed from commercial leases, research clusters, and the supercomputing ecosystem around Oak Ridge, Tennessee. Every pool is load-tested on the models it serves, and the measured figures drive capacity planning. Sustainable.
Three-site routing. Requests are routed across three physically separate sites, and the relay steers around capacity that is unhealthy or overloaded at any one of them. Resilient.
Explicit refusals. Every rejection carries a source, a machine-readable code, whether the request was charged, and when to retry, so that an application can tell a plan limit from fleet capacity from a platform fault. Accessible.
Usage-value plans with transparent windows. Subscriptions are denominated in usage value over a short burst window and a longer sustained window, with reserve-then-settle metering so that limits mean what they say. Budget state is returned in response headers on every request. Prepaid credit is available for pay-as-you-go use. Accessible.
Energy efficiency, measured. A completion on the network uses roughly a third of the energy of a typical industry equivalent, on hardware the industry had retired. The fleet uses no evaporative cooling. Sustainable.
Renewable power, phase one. On-site solar supplies a measured share of completion power, reported as a percentage of what actually serves requests rather than as an offset. Sustainable.
Zero retention. Request contents are not stored after a response is served, and the network's logging and vendor configuration are reviewed to confirm it. Accessible.
Operational monitoring and automatic capacity gating. Per-node power, temperature, utilization, latency, and error rate are tracked continuously. Scheduled checks alert a person when a rule is breached, a morning report summarizes health, capacity, security, and subscription load, and new subscriptions are gated against measured capacity so that growth never outruns the hardware. Resilient.
In active development
Per-request energy accounting in the dashboard. The fleet already measures node power, and usage metering already records prompt, cached, and generated tokens per request. The attribution layer that combines them into a per-request energy figure, with published allocation rules for idle power and shared hardware, is being built so that energy appears in the dashboard next to cost. Sustainable.
Renewable power, phase two. Expanding on-site generation and the measured solar share of completion power, with the goal of scheduling batch workloads into the hours when local generation is highest. Sustainable.
Expanded model availability. Adding capacity for the architectures developers are asking for most, with particular attention to sparse mixture-of-experts models that are cheap per token on the hardware the network runs, and to larger-context configurations for agent workloads. Every addition goes through the same conformance gate before it is listed. Accessible.
Improved agent support. Streaming and tool-call behavior validated against the agent frameworks developers actually use, and clearer guidance on structuring prompts so that stable prefixes are cached. The budget and refusal work completed this year is the foundation; this phase is about making agents feel native on the network. Accessible.
Availability through third-party marketplaces. Making the network's models available to developers through the model marketplaces they already use, with the same prices, limits, and refusal semantics as direct access, and with capacity reserved so that marketplace traffic does not crowd out direct subscribers. Accessible.
Municipal installation planning. Cobble is in active discussions with communities in Western Kentucky and Eastern Kentucky about building locally owned inference infrastructure connected to the network. This phase is about site assessment, power, and the operating model, not yet construction. Resilient.
Longer-term ambitions
Community-owned installations in production. Turning the Kentucky discussions, and others like them, into running facilities: locally owned, serving their regions, connected to the network, and earning for the communities that operate them. This depends on partnerships and capital that are not yet in place, and the timeline belongs to the communities as much as to Cobble. Resilient.
A shared foundation for other operators. Packaging the relay, routing, metering, refusal semantics, and energy accounting so that a new installation can join the network without rebuilding any of it. The pieces exist in production; making them something another operator can adopt is substantial work. Resilient.
Energy budgets alongside spending budgets. Once per-request energy is in the dashboard, letting an application carry an energy target the way it carries a spending target, with trends and the information needed to meet it. Sustainable.
Waste-heat reuse. Small installations produce heat at roughly the scale a neighboring building, greenhouse, or district loop could use. This is an integration that only becomes possible at small scale, and it is an ambition rather than a plan. Sustainable.
Renewable power at majority share. Growing on-site and local generation until it supplies the majority of completion power across the network, measured the same way it is today. Sustainable.
How the pieces connect
It would be easy to read the list above as unrelated projects. It is not. Each one removes a reason that inference has to be centralized.
Reclaimed hardware and measured efficiency remove the need for the newest silicon. Local renewable power removes the need for a hundred-megawatt grid connection. Three-site routing and explicit refusals remove the single point of failure. Usage-value plans and capacity gating remove the mismatch between what is sold and what exists. Conformance testing and agent support remove the gap between open-weight models and the proprietary ones developers are used to. Energy accounting removes the excuse that environmental cost cannot be measured. And the municipal installations, when they arrive, remove the assumption that the infrastructure has to belong to someone far away.
That is the mission in one sentence: inference that is sustainable because it runs efficiently on what already exists, resilient because it is distributed, and accessible because it is open, honest about its limits, and owned closer to the people who use it.
Reading this roadmap
Items in the first section are things a customer can rely on today, and the network's documentation describes them in detail. Items in the second have engineering time committed and will move to the first section as they ship; they are not promises of a date. Items in the third are directions, and they are listed so that partners and customers can see where the network intends to go, not so that anyone plans around them.
This page will be updated as items move between sections. Where something slips or is abandoned, that will be said here too.
The Cobble Team